Abstract
The objective quantification of pediatric craniofacial anomalies is essential for diagnosis of pathology and treatment planning. However, clinical evaluation still relies on subjective assessment or metrics with low sensitivity because existing craniofacial anatomy models have been mostly developed for adult applications using methods that typically fail to capture rapid anatomical changes during pediatric growth given the unavailability of large normative pediatric datasets. Moreover, they normally require extensive pre-processing and manual annotations to create standardized anatomical surface meshes and often impose strong statistical assumptions. We introduce a novel geometric learning model that operates directly on diverse anatomical surface representations of pediatric patients with pathology and does not require preprocessing. Our model infers the normative distributions in the population by disentangling anatomical representations from pathology, age and sex information. First, it learns a probabilistic mapping between input surfaces and a uniform, geometrically consistent latent space. Then, a new variational-adversarial training framework removes pathology, age and sex information to learn population-wide statistical distributions, which enables either generating personalized normative references to characterize pathology or realistic synthetic data. Trained on 2,696 3D photograms with diverse craniofacial pathologies and evaluated on normative anatomical regions, our model achieved a point-to-surface distance of 1.50±0.72 mm. Moreover, our personalized normative references allowed identifying phenotypes of cranial pathology with 96% accuracy. Finally, synthetically generated data showed no significant volumetric differences compared to real data during pediatric development (p=0.13). Our results show that our method can generate personalized normative anatomical references and synthesize realistic pediatric datasets.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/ShapeMI_028.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=xOO16IfNmn
BibTex
@InProceedings{NagJos_Disentangled_MICCAISAT2026,
author = { Nagel, Joseph AND Cruz-Guerrero, Ines A. AND French, Brooke AND Porras, Antonio R.},
title = { { Disentangled Geometric Variational Learning of Normative Pediatric Development from Anatomical Surface Representations of Patients with Pathology } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17259},
month = {pending},
page = {pending}
}
